Searches are increasingly answered by systems that show an answer rather than a list of links: AI Overviews in Google Search, ChatGPT with web search, Perplexity, Copilot, Gemini. That raises a new question for companies: are we mentioned in these answers, and if so, correctly and with a source? This article describes how to measure it without relying on tool promises.
Why Search Console is not enough
Search Console shows impressions and clicks in Google Search. Appearances in AI Overviews are included in the normal web data without being reported separately. Answers from ChatGPT, Perplexity or Copilot do not appear at all. Judging visibility through Search Console alone misses a growing part of search. This applies to this website too: for “generative engine optimization” our own glossary page appears in Search Console with several hundred impressions but beyond position 60. What AI systems answer to that question, Search Console does not show.
The metrics
- Mention rate: in how many answers to a fixed set of questions is the company or product mentioned?
- Citation rate: how often is one of your own pages linked as a source?
- Prominence: is the mention at the start of the answer or in a long list?
- Accuracy: are statements about services, prices, locations and products correct?
- Referral traffic: how many visits come from chatgpt.com, perplexity.ai and similar sources?
- Crawler access: which AI crawlers fetch which pages, visible in server or hosting logs?
The first four are measured through a prompt panel, the last two through analytics and server data. No single metric is enough. A high mention rate with wrong statements is a risk, not an achievement.
Building a prompt panel
A panel consists of 30 to 60 questions real customers would ask, in four groups: brand questions, category questions (“Which bid management software providers are there in Switzerland?”), problem questions (“How do I prepare for the Swiss cyberattack reporting obligation?”) and comparison questions. Questions are asked in the languages the company sells in, with regional context where relevant, and the wording stays the same month to month, otherwise results are not comparable.
Run it monthly in every relevant system, without login or with a neutral account, so personalisation does not distort results. Because AI answers vary, each question is asked several times and the result recorded as a share rather than yes or no. A simple spreadsheet with question, system, date, mention, source and notes is enough.
From measurement to action
The analysis typically shows three patterns. First, gaps: the company is not mentioned for category questions although it has relevant pages. Self-contained pages that answer the question directly, clear entity information about the company and its products (entity SEO) and mentions on third-party sites that AI systems use as sources help here. Second, errors: the answer names outdated offerings or wrong locations. Consistent information on the website, business profile and directories helps. Third, missing citations: the company is mentioned but not linked. Pages with precise, citable statements and sources help.
Technically, ensure AI crawlers may and can fetch the pages. Whether to allow or block them is a business decision weighed in our decision guide on AI crawlers. Because several AI search systems rely on their own or third-party indexes, maintaining Bing Webmaster Tools and submitting changed pages via IndexNow is worthwhile alongside Search Console.
Limits of measurement
AI answers are not deterministic, systems change without notice and providers publish no official data. A panel therefore measures trends, not exact values. Tools promising “AI visibility” to one decimal place sell false precision. An honest panel uses fixed questions, several runs and an openly documented method.
Next steps
The technical foundations are covered by the GEO and AEO readiness checklist. We offer an ongoing panel with analysis and an action plan as AI visibility monitoring. The concepts are explained in the glossary entry on generative engine optimization.